Simulate YOLOv5 Segmentation Inference

tip

This document demonstrates how to use rknn-toolkit2 on an x86 PC to perform simulation inference of the YOLOv5 segmentation model without a motherboard. For the required environment setup, please refer to RKNN Installation.

Prepare the Model

This example uses a pre-trained ONNX model from the rknn_model_zoo as a case study, converting the model and running simulation inference on the PC.

• If you are using Conda, first activate the rknn Conda environment:

  X86 Linux PC

  conda activate rknn

• Download the yolov5s-seg.onnx model:

  X86 Linux PC

  cd rknn_model_zoo/examples/yolov5_seg/model
  # Download the pre-trained yolov5s-seg.onnx model
  bash download_model.sh

  tip

  If you encounter network issues, you can visit this page to manually download the model and place it in the corresponding folder.

• Rename the ONNX model file suffix to .rknn (only for PC-side result simulation):

  X86 Linux PC

  cp yolov5s-seg.onnx yolov5s-seg.rknn

• (Optional) Use rknn-toolkit2 to convert the model to yolov5s-seg.rknn:

  X86 Linux PC

  cd rknn_model_zoo/examples/yolov5_seg/python
  python3 convert.py <onnx_model> <TARGET_PLATFORM> <dtype> <output_rknn_path>
  # python3 convert.py ../model/yolov5s-seg.onnx rk3588 i8 ../model/yolov5s-seg.rknn

  Parameter explanation:

  • <onnx_model>: Path to the ONNX model.
  • <TARGET_PLATFORM>: Name of the NPU platform. Options: rk3562, rk3566, rk3568, rk3576, rk3588, rk1808, rv1109, rv1126.
  • <dtype>: Choose i8 or fp. i8 is for INT8 quantization; fp is for FP16 quantization. The default is i8.
  • <output_rknn_path>: Path to save the RKNN model. By default it is saved in the same directory as the ONNX model with the filename yolov5s-seg.rknn.
  tip

  For RK358X users, set TARGET_PLATFORM to rk3588.

ONNX Model Inference on PC

X86 Linux PC

python3 yolov5_seg.py --model_path ../model/yolov5s-seg.onnx --img_show

Result

person @ (212 242 285 510) 0.871
person @ (111 240 223 535) 0.850
person @ (472 233 559 519) 0.831
bus  @ (97 134 549 458) 0.799
person @ (80 328 125 517) 0.470

RKNN Model Simulation Inference on PC

• Install the required dependencies with pip3:

  X86 Linux PC

  pip3 install torchvision==0.19.0 pycocotools

• Run the simulation inference script:

  • Modify rknn_model_zoo/py_utils/rknn_executor.py to the following code (be sure to back up the original file):

    Python Code

    from rknn.api import RKNN
    
    class RKNN_model_container():
        def __init__(self, model_path, target=None, device_id=None) -> None:
            rknn = RKNN()
            DATASET_PATH = '../../../datasets/COCO/coco_subset_20.txt'
            onnx_model = model_path[:-4] + 'onnx'
            rknn.config(mean_values=[[0, 0, 0]], std_values=[[255, 255, 255]], target_platform=target)
            rknn.load_onnx(model=onnx_model)
            rknn.build(do_quantization=True, dataset=DATASET_PATH)
            rknn.init_runtime()
            self.rknn = rknn
    
        def run(self, inputs):
            if isinstance(inputs, list) or isinstance(inputs, tuple):
                pass
            else:
                inputs = [inputs]
    
            result = self.rknn.inference(inputs=inputs)
            return result
    
        def release(self):
            self.rknn.release()
            self.rknn = None

  • Run the simulation inference script:

    X86 Linux PC

    python3 yolov5_seg.py --target <TARGET_PLATFORM> --model_path <RKNN_MODEL_PATH> --img_show
    # python3 yolov5_seg.py --target rk3588 --model_path ../model/yolov5s-seg.rknn --img_show

    --target: Name of the NPU platform to simulate. Options: rk3562, rk3566, rk3568, rk3576, rk3588, rk1808, rv1109, rv1126.

    --model_path: Path to the RKNN model to be simulated.

    Results

    person @ (213 239 284 516) 0.882
    person @ (109 240 224 535) 0.869
    person @ (473 231 560 523) 0.845
    bus  @ (97 136 548 459) 0.821
    person @ (80 328 124 519) 0.499

  • Simulation inference result (the simulator only simulates NPU computation; actual performance and accuracy are determined by inference on the target board):